Modeling partly conditional means with longitudinal data

We propose a general modeling approach to longitudinal data that is a hybrid of the marginal regression models of Zeger and Liang and of the classical transition models such as used in time series analyses. Rather than conditioning at time t only on covariate values, as is typical with the marginal approach, or on the entire history of the process up to t, as is typical with the transition model approach, we suggest models that condition on a subset of the process history. Estimation proceeds using generalized estimating equation methodology but with the restriction that the working covariance matrix is diagonal. The proposed regression models share common features with Cox regression models for failure time data in that they are composed of a nuisance baseline function of time and a simple parametric function of the covariates. Two illustrative examples are presented. © 1997 Taylor & Francis Group, LLC.